# AgentSeal — Security Toolkit for Agents & MCP

> AgentSeal is an FSL-1.1 (future Apache-2.0) toolkit that scans skills and MCP configs, monitors supply-chain risks, and tests prompt-injection resistance.

## Install

Save the content below to `.claude/skills/` or append to your `CLAUDE.md`:

## Quick Use

```bash
pip install agentseal
agentseal guard            # offline scan (no API key)
agentseal scan --file ./prompt.txt --model ollama/llama3.1:8b
agentseal scan-mcp --server npx @modelcontextprotocol/server-filesystem /tmp
```

## Intro

AgentSeal is an FSL-1.1 (future Apache-2.0) toolkit that scans skills and MCP configs, monitors supply-chain risks, and tests prompt-injection resistance.

**Best for:** teams hardening agent configs before production and continuously monitoring prompt/MCP attack surfaces

**Works with:** Python or Node installs, Ollama for local inference, MCP servers (stdio/SSE) for audits

**Setup time:** 5-10 minutes

### Key facts (verified)

- GitHub: 254 stars · 38 forks · pushed 2026-04-29.
- License: FSL-1.1-Apache-2.0 · owner avatar + repo URL verified via GitHub API.
- README-verified entrypoint: `pip install agentseal # or: npm install agentseal`.

## Main

- Use `guard` as a preflight: scan skills and MCP configs for risky patterns before you ship a new agent build.

- Use `scan` to measure prompt robustness over time—treat it as a regression test suite for system prompts.

- Audit MCP servers you rely on: tool descriptions are part of your trust boundary, so poisoning detection matters.

### Source-backed notes

- README Quick Start includes `agentseal guard` for offline scanning with no API key.
- README lists commands like `scan` (prompt probes) and `scan-mcp` (audit live MCP tool descriptions).
- README mentions baseline tracking with SHA-256 hashes to detect changes since last scan.

### FAQ

- **Does it work offline?**: Yes—README says `guard`, `shield`, and `scan-mcp` can work offline.
- **Do I need a model for scanning prompts?**: For `scan`, yes; use Ollama locally or a cloud model.
- **Is it open source?**: License is FSL-1.1 (future Apache-2.0); review LICENSE for usage terms.

## Source & Thanks

> Source: https://github.com/getagentseal/agentseal
> License: FSL-1.1-Apache-2.0
> GitHub stars: 254 · forks: 38

---

<!-- ZH -->

## Quick Use

```bash
pip install agentseal
agentseal guard            # offline scan (no API key)
agentseal scan --file ./prompt.txt --model ollama/llama3.1:8b
agentseal scan-mcp --server npx @modelcontextprotocol/server-filesystem /tmp
```

## Intro

AgentSeal 是 FSL-1.1（未来 Apache-2.0）许可的安全工具箱，可扫描危险技能与 MCP 配置、监控供应链风险，并做提示词注入（prompt injection）抗性测试，适合把 agent 上线前的安全检查流程化。

**Best for:** 要把 agent 配置上线前做安全体检，并持续监控 prompt/MCP 攻击面的团队

**Works with:** Python 或 Node 安装方式、Ollama 本地推理、可审计的 MCP server（stdio/SSE）

**Setup time:** 5-10 minutes

### Key facts (verified)

- GitHub：254 stars · 38 forks；最近更新 2026-04-29。
- 许可证：FSL-1.1-Apache-2.0；作者头像与仓库链接均已通过 GitHub API 复核。
- README 中核对过的入口命令：`pip install agentseal # or: npm install agentseal`。

## Main

- 把 `guard` 当上线前检查：扫描技能文件与 MCP 配置里的高风险模式。

- 用 `scan` 做提示词鲁棒性回归：把系统 prompt 当测试对象长期维护指标。

- 对依赖的 MCP server 做审计：工具描述本身也是信任边界，投毒检测很关键。

### Source-backed notes

- README 的 Quick Start 包含 `agentseal guard`，可离线扫描且无需 API key。
- README 提供 `scan`（提示词探测）与 `scan-mcp`（审计 MCP 工具描述）等命令。
- README 提到用 SHA-256 做基线追踪，以发现配置被篡改或供应链变更。

### FAQ

- **可以离线用吗？**：可以。README 说明 `guard`/`shield`/`scan-mcp` 可离线运行。
- **扫 prompt 一定要模型吗？**：`scan` 需要；可用 Ollama 本地或云端模型。
- **它算开源吗？**：许可为 FSL-1.1（未来 Apache-2.0）；具体以 LICENSE 条款为准。

## Source & Thanks

> Source: https://github.com/getagentseal/agentseal
> License: FSL-1.1-Apache-2.0
> GitHub stars: 254 · forks: 38


---
Source: https://tokrepo.com/en/workflows/agentseal-security-toolkit-for-agents-mcp
Author: Script Depot